ADM-201 dump PMP dumps pdf SSCP exam materials CBAP exam sample questions

AI学习3D面部重构 – 译学馆
未登录,请登录后再发表信息
最新评论 (0)
播放视频

AI学习3D面部重构

AI Learns 3D Face Reconstruction | Two Minute Papers #198

亲爱的学者们
Dear Fellow Scholars,
欢迎来到Károly Zsolnai-Fehér的两分钟论文
this is Two Minute Papers with Károly Zsolnai-Fehér.
现今面部识别逐渐成为一个热门话题
Now that facial recognition is becoming more and more of a hot topic,
那么我们就来聊聊3D人脸重建吧!
let’s talk a bit about 3D face reconstruction!
在这个问题中 我们要有一张2d照片
This is a problem where we have a 2D input photograph,
或者是某人的一段视频
or a video of a person,
目的是用它来创建出一个3D几何模型
and the goal is to create a piece of 3D geometry from it.
为了实现这个目标 以前通常
To accomplish this, previous works often required
需要面部高度对称
a combination of proper alignment of the face,
多张照片以及照片密集对应
multiple photographs and dense correspondences,
对于分辨这些照片相同区域
which is a fancy name for additional data
的附加数据来说 密集对应 是一个奇特的名字
that identifies the same regions across these photographs.
但这个新的构想
But this new formulation is the holy grail
是这个问题所有可能版本的“圣杯”(焦点)
of all possible versions of this problem,
因为它只需要一张2D照片
because it requires nothing else but one 2D photograph.
这项工作选择的算法是卷积神经网络
The weapon of choice for this work was a Convolutional Neural Network,
并且常用算法的数据集已经被精简的不能再简单了
and the dataset the algorithm was trained on couldn’t be simpler:
它只有一个大的2D输入图像数据库
it was given a large database of 2D input image
搭配3D几何输出模型
and 3D output geometry pairs.
这意味着神经网络能考虑到大量这类搭配
This means that the neural network can look at a lot of these pairs
并且了解 如何用这些输入照片来映射出3D几何模型
and learn how these input photographs are mapped to 3D geometry.
正如你所看到的那样 这个结果是绝对震撼的
And as you can see, the results are absolutely insane,
尤其是考虑到它能够作用于任意的面部位置
especially given the fact that it works for arbitrary face positions
许多不同的表情 即使面部有被遮掩
and many different expressions, and even with occlusions.
然而 这并不是传统的卷积神经网络算法
However, this is not your classical Convolutional Neural Network,
因为正如我们曾提到的 它的输入是2D 输出是3D
because as we mentioned, the input is 2D and the output is 3D.
所以问题立即被提出
So the question immediately arises:
应该在输出中采用哪种数据结构
what kind of data structure should be used for the output?
作者选择了一种3D立体像素排列
The authors went for a 3D voxel array,
本质上 是一种方块 我们用以建立
which is essentially a cube in which we build up
如同小的 同样的“乐高”碎片组成的面部
the face from small, identical Lego pieces.
这种表现类似我的世界游戏中的地势
This representation is similar to the terrain in the game Minecraft,
只有这些区域的分辨率是好的
only the resolution of these blocks is finer.
猜测这些立体像素的排列的方法
The process of guessing how these voxel arrays should look
应该注意基本的输入照片
based the input photograph is referred to
这个过程参考了群落研究中的体积复原
in the research community as volumetric regression.
这是3D技术的本质
This is what this work is about.
并且现在已经成为最棒的部分
And now comes the best part!
我们也可以通过一份在网上能得到的演示
An online demo is also available
尝试处理一些准备好的图片
where we can either try some prepared images,
或者 我们也可以上传我们自己的图片
or, we can also upload our own.
所以当我运作我的实验时
So while I run my own experiments,
不要忘记我的优秀的素材
don’t leave me out of the good stuff
并确认你在评论部分展示你的结果
and make sure you post your results in the comments section!
你也可以获得这些原始代码 当你跟着修补人员离开这儿
The source code is also available for you fellow tinkerers out there.
3D技术的局限性包括 不能检测
The limitations of this technique includes the inability of detecting expressions that
那些人们在训练装置中看到的偏差很大的表情
are very far away from the ones seen in the training set,
或者你可以在录像中看到的表情
and as you can see in the videos,
片段的连贯性也需要改进
temporal coherence could also use some help.
这意味着当你输入一段录像时
This means that if we have video input,
成品会有一些微小的差异 存在于每个框架中
the reconstruction has some tiny differences in each frame.
在不久的未来 可能一个周期性神经网络
Maybe a Recurrent Neural Network,
像是 长短期记忆变体
like some variant of Long Short Term Memory
会解决这个问题
could address this in the near future.
然而 神经网络将更加机警更加资源集中利用用来恰当地训练
However, netr more resources properly.
非常兴奋的看到这些解决方法逐渐形成
Very excited to see how these solutions evolve,
当然 两分钟论文将一直在这里陪伴着你
and of course, Two Minute Papers is going to be here for you
讨论那些令人惊叹的即将到来的作品
Very excited to see how these solutions evolve,
感谢你们的观看和支持
Thanks for watching and for your generous support,
我们下次再见
and I’ll see you next time!

发表评论

译制信息
视频概述

介绍了3D面部重构的理论基础

听录译者

收集自网络

翻译译者

Taoyasa

审核员

审核员 W

视频来源

https://www.youtube.com/watch?v=9BOdng9MpzU

相关推荐